Somewhere on a spreadsheet inside every large cloud provider sits a number called Remaining Performance Obligations — RPO, in the acronym Wall Street has learned to trade on. It is not revenue. It is not cash. It is a promise: a customer has signed a contract to pay for compute the provider has not yet delivered and, in the more aggressive cases, has not yet built. In March 2026, Oracle disclosed an RPO of $553 billion and its stock added roughly a quarter-trillion dollars of market value in a single session — in a quarter where its actual revenue and earnings missed consensus.8 Part 1 followed the shell; Part 2 followed the silicon. This part follows the token one rung further up, into the layer that rents you all of it on a contract — and finds that the contract itself has become the asset the market prices, whether or not anyone can say when, or whether, it gets paid.
Part 3 of The Stack. We are one rung up from the silicon. Same two questions, asked of the cloud/PaaS layer: who books the equity, and who eats the subsidized loss? Here the equity story is roughly two trillion dollars of disclosed backlog moving stock prices on announcement; the loss, when it lands, lands unevenly — on a credit-rating scale you can read directly in basis points.
/ 01The Two-Trillion-Dollar Receivable
Start with the plain-English definition, because the acronym does a lot of work to obscure it. Per Oracle's own accounting language, RPO is “deferred revenues; invoices issued but uncollected; and amounts to be invoiced in future periods”8 — a blend of cash the company may already be sitting on and a legally binding promise to pay that has not yet been billed, let alone collected. Sherwood News' own worked example is the cleanest: an earlier, smaller Oracle–OpenAI contract, worth $30 billion a year, will not become recognized revenue until fiscal 2028, and two-thirds of a prior $137.8 billion RPO balance would not be recognized as revenue for at least twelve months from disclosure.8 An RPO, in other words, is a receivable booked today against a counterparty's ability to pay years from now — and, per JPM's own reporting, roughly half of the industry's RPO is booked against two counterparties, OpenAI and Anthropic, whose own revenue does not yet cover what they've promised.24
This is not a single company's accounting quirk. Pull the number directly from each provider's own most recent filing or earnings release and it sums to real money: Oracle $553 billion (fiscal Q3 2026, period ended February 28, 2026);1 Microsoft $633 billion total company, $627 billion of it commercial (fiscal Q3, ended March 31, 2026);2 Alphabet $467.6 billion total, $462.3 billion of it Google Cloud (Q1 2026, ended March 31, 2026);6 Amazon/AWS $364 billion (Q1 2026, ended March 31, 2026);3 CoreWeave $98.8 billion (Q1 2026, ended March 31, 2026);4 Nebius roughly $50 billion contracted (as of its May 13, 2026 earnings).5 Add them and you land at roughly $2.17 trillion — a bottoms-up tally built independently from six companies' own disclosures, and it corroborates, almost exactly, JPM's own headline figure of “around $2 trillion…as of Q1,” sourced separately to The Information.24 Two independent methods, two SEC-filing-adjacent sources, one number in the same neighborhood. Treat the total as real.
Here is the Keene move, stated as plainly as the evidence allows: an RPO is a receivable booked against a counterparty whose revenue does not yet exist. The labs' spend is the clouds' backlog is the labs' revenue — and the market, watching Oracle rally a quarter-trillion dollars on the strength of a number that is that circularity, is pricing the promise, not the delivery.
RPO is not vapor. It is a real, SEC-disclosed, legally binding number, and for most of the providers in this table, the overwhelming majority of it is owed by a genuinely diversified base of paying enterprise customers with current revenue — Microsoft's own filing shows commercial RPO excluding OpenAI still grew 26% year over year, which is a healthy business by any measure.2 Buffett would tell you a signed, enforceable contract is worth something even before the cash arrives. The honest claim here is not “RPO is a lie.” It is that RPO quality varies enormously by provider — and the market, as of this writing, is not yet pricing that variance. Oracle's rally on a quarter it missed is the clearest evidence of that.
Microsoft's and Alphabet's totals are company-wide RPO — they include non-cloud commercial commitments (Microsoft 365, Google Search and Workspace), not pure cloud-infrastructure backlog, so the true PaaS-only total is somewhat smaller than $2.17 trillion.26 And Bank of America's widely cited estimate that OpenAI is roughly 54% ($300 billion) of Oracle's $553 billion RPO is a sell-side analyst's estimate, reproduced via a secondary aggregator — Oracle itself does not break out RPO by named customer in its earnings materials.13
The cloud books it, and the stock moves
The commitment becomes RPO the instant it's signed — a headline number that moves the provider's share price on announcement, independent of whether a dollar has actually been collected.8
The backlog becomes collateral
A bankable RPO print lets the provider borrow against it — Oracle's $25 billion investment-grade bond sale, six times oversubscribed, priced weeks after an RPO disclosure.10 The circle closes without a dollar of revenue from an unrelated third party.
/ 02The Landlords of the Boom
Per Synergy Research's Q1 2026 count, the “big three” — AWS (28%), Azure (21%), Google Cloud (14%) — hold 63% of a $129-billion-per-quarter enterprise cloud infrastructure market, an annualized run rate over half a trillion dollars, growing 35% year over year for a ninth consecutive quarter of accelerating growth.18 Tellingly, Synergy's tier-two growth list now includes OpenAI and Anthropic themselves — the frontier labs are beginning to show up as infrastructure operators in market-share data, not just as customers of it.18
AWS / Bedrock. Amazon's cloud segment posted Q1 2026 net sales of $37.6 billion and operating income of $14.2 billion — a margin above 37%, still the profit engine of Amazon overall — while Bedrock, AWS's managed model-serving layer, sells a curated catalog (Claude, Titan, Llama, Mistral, Cohere, DeepSeek) bundled with exclusive Claude access and AWS's own enterprise data-residency commitments.315
Azure / Foundry. Azure and other cloud services revenue grew 40% year over year in fiscal Q3 2026; Microsoft Cloud overall (Azure, M365 Commercial cloud, Dynamics 365, LinkedIn commercial) hit $54.5 billion, up 29%.2 Foundry is the only PaaS offering the full GPT-5/o3 family under enterprise data-protection terms — a genuine exclusivity moat — and added DeepSeek V4 to its catalog in May 2026.19 Microsoft's OpenAI relationship is now a 26.79%-fully-diluted equity stake, and its own 10-Q discloses $5.9 billion of net gains from that investment over nine months, driven “primarily” by a dilution gain from OpenAI's October 2025 for-profit recapitalization — an accounting gain generated by OpenAI restructuring itself, not by OpenAI paying Microsoft anything.220 Microsoft's own non-GAAP “adjusted” EPS excludes these swings — a useful, company-authored admission that the number is volatile and not operational.2
Google Cloud / Vertex. Google Cloud's backlog more than doubled sequentially, from $240 billion at the end of 2025 to $462.3 billion in Q1 2026 — a jump Alphabet itself attributes partly to “the inclusion of TPU hardware sales,” meaning Google is now booking multi-gigawatt TPU hardware sales into cloud backlog, not just consumption-based API revenue.6 The same 10-Q discloses a $40.0 billion commitment to an unnamed private company and the $29.5 billion closing of the Wiz cybersecurity acquisition, folded into the Cloud segment.6
Oracle Cloud. Cloud revenue was $8.9 billion in fiscal Q3 2026, up 44%, against total company revenue of $17.2 billion — cloud is roughly half of Oracle and its only meaningfully growing segment.1 Part 1 already tallied the debt Oracle is taking on to fund the buildout underneath the Stargate commitment;9 what's new here is the price of that debt. In February 2026 Oracle priced a $25 billion, eight-tranche bond offering at coupons of 4.55% to 6.85%, with its ten-year tranche pricing roughly 145 basis points over Treasuries against an average ~95 basis points for BBB peers — the market openly charging Oracle a premium above its own rating tier, even as the deal went six times oversubscribed.10 Oracle's capex guidance has been revised upward three times in roughly a year — $25 billion, then $35 billion, then $50 billion — against S&P's own forecast that capex could clear $60 billion in fiscal 2028.121
CoreWeave, Nebius, and the neoclouds. CoreWeave's revenue grew 112% year over year to $2.08 billion in Q1 2026, even as its stock traded around $86.72 on July 1, down roughly 48% from its 52-week high — the market repricing risk, not revenue.7 Nebius posted Q1 2026 revenue of $399 million, up 684% year over year, with a roughly $50 billion contracted backlog anchored by a $27 billion Meta deal and up to $19.4 billion from Microsoft — meaning two counterparties account for something close to 93% of Nebius's headline backlog.5 Synergy separately sizes the broader neocloud tier — Lambda, Crusoe, Vast Data, Core Scientific, TensorWave — at $35.2 billion of 2026 revenue, projected toward $236.5 billion by 2031; that is a market-research forecast, not a reported figure, and should be read as directional texture only.22
Now the falling-margin half of the picture, and Amazon supplies the cleanest, most directly sourced case in the entire brief. Amazon's own 10-Q free-cash-flow reconciliation shows trailing-twelve-month operating cash flow rising from $113.9 billion to $148.5 billion between the quarters ended March 2025 and March 2026 — a genuinely growing business — while trailing capex climbed from $88.0 billion to roughly $151 billion over the same span. The result: trailing-twelve-month free cash flow collapsed from $25.9 billion to $1.2 billion, a company-reported, SEC-disclosed decline of about 95%.3 This is not operational failure. It is arithmetic: capex is outrunning operating cash flow, and it is happening at the company with the deepest, most diversified balance sheet in the group. Combined 2026 capex across the four largest hyperscalers is guided toward roughly $725 billion, up about 77% from roughly $410 billion in 2025 — an analyst aggregate of individual company guidance, not a sum any one company discloses itself, but each input is consistent with that company's own stated numbers.23
Who is actually profitable versus who is booking backlog is, at this point, a legible split. Microsoft, Amazon, and Google can absorb the RPO circularity inside a much larger, genuinely profitable core business — Azure and AWS both post real operating margins on real consumption revenue, with or without a single frontier-lab dollar. Oracle's cloud transformation is substantially the OpenAI bet, funded with negative-outlook investment-grade debt priced at a visible premium. CoreWeave and Nebius sit a rung further out, funding the identical underlying exposure — frontier-lab compute demand — with debt that costs roughly double what Oracle pays.
/ 03The Relocated Toll Booth
Part 2 documented open-weight models — DeepSeek, Qwen, Llama, Mistral — closing the quality gap at a fraction of the price. The intuitive conclusion is that this dissolves hyperscaler lock-in: if the model is free and the weights are open, why pay AWS or Microsoft anything at all? The evidence in this layer says otherwise, and JPM's own language is exact: “open models do not necessarily imply worse hyperscaler enterprise-grade guarantees or adoption complexity if you opt for turnkey hyperscaler inference services.”14
An enterprise that wants to run DeepSeek V4 with SOC2 or HIPAA-grade guarantees, EU data residency, customer-managed keys, and a contractual SLA does not typically self-host it. It accesses the open model through Bedrock or Foundry — paying the hyperscaler's PaaS margin regardless of whose weights sit underneath.1415 DeepSeek V4's own arrival in the Microsoft Foundry catalog in May 2026 is the concrete instance: the open, free-to-download model becomes, inside Foundry, a paid, metered, hyperscaler-branded product with a compliance wrapper around it.19 The genuinely disruptive path — do-it-yourself infrastructure, self-hosting open weights without a hyperscaler intermediary — is, in JPM's own assessment, “unlikely to be adopted broadly since it requires a more sophisticated hands-on approach.”14 Most enterprises will not build that muscle. They will rent the guarantee instead.
This is the toll booth moving, not disappearing. The commoditization of the model is real — Part 2 wasn't wrong about that. But the thing an enterprise actually pays for was never really the weights; it was the promise that its data stays where the law says it has to stay, backed by a company with the balance sheet to be sued and to lose. That promise is the hyperscaler's to grant, to price, or in principle to withdraw — and it is exactly as true underneath a $60,000-a-seat frontier subscription as it is underneath a model you could download for free. Hold this thread; it is the load-bearing idea behind the data-sovereignty argument Part 6 makes about who actually escapes the ring.
/ 04Counterparty Risk, in the Filings
The strongest evidence in this entire layer is not analyst commentary. It is two companies' own regulatory language, independently arriving at the Keene framing.
Moody's rates Oracle Baa2 with a negative outlook. Its stated rationale is not abstract leverage anxiety — it is named, specific, and about exactly one thing:
“High reliance on revenue from a single counterparty.” — Moody's Ratings, on Oracle's negative outlook, 2026
That is a credit-rating agency writing, in its own institutional language, the same sentence this series has been building toward since Part 1: the equity holder collects a headline number; the risk sits with whoever has to be paid back if the counterparty can't perform. S&P's independent read arrives at the same place from the debt side — it affirmed Oracle at BBB, negative outlook, forecasting the free-operating-cash-flow deficit widening well past Oracle's own $10 billion guidance and leverage potentially clearing 4x in fiscal 2027–28, a threshold S&P had to loosen from its own prior trigger of 3.5x specifically to accommodate Oracle's trajectory without a downgrade today.21 Both agencies are affirming an investment-grade rating that rests substantially on one customer's ability to eventually pay $300 billion it does not yet have.
CoreWeave's own 10-Q makes the same admission in plainer language, and it is worth quoting directly rather than paraphrasing, because a company rarely states its own tail risk this cleanly:
“As we expand our customer base, we may become further subject to counterparty credit risk…We do not currently maintain credit insurance to insure against customer credit risk…some of our customers may be highly leveraged.” — CoreWeave, Inc., Form 10-Q, quarter ended March 31, 2026
That is a filed, legally reviewed risk factor, not a critic's inference. CoreWeave's own numbers show why it's there: its largest customer, “Customer A,” was 45% of Q1 2026 revenue — down from 72% a year earlier, which is genuine diversification, but still acute concentration by any normal standard — while OpenAI alone has committed up to $18.4 billion combined across two order forms running through 2031.4 CoreWeave funds its buildout with roughly $20 billion of debt across term-loan and note facilities carrying effective interest rates of 6% to 15% — materially above the 4.55%–6.85% Oracle just paid on its own investment-grade issue.410 The credit-risk premium here is not a metaphor. It is priced, filed, and visible in basis points, and it is the neocloud tier — not the hyperscalers — absorbing it.
“The rating agency and the risk-factor footnote are saying the same sentence. One just gets read by more people.”
It would be dishonest to let this section imply every cloud provider carries Oracle's or CoreWeave's exposure. They don't. Microsoft's ex-OpenAI commercial RPO still grew 26% year over year — a real, diversified enterprise business underneath the headline number.2 The counterparty-concentration critique bites hardest specifically at Oracle and, one rung down the credit ladder, at CoreWeave and Nebius — not at the sector as a whole. The honest synthesis is not “cloud backlog is fake.” It's that backlog quality varies by an order of magnitude within this single layer, and the credit-rating gap between Oracle's 4.55%–6.85% and CoreWeave's 6%–15% is the market's own receipt for that variance.
/ 05Sovereign Compute, Two Ways
Sovereign compute stopped being a slide-deck concept sometime in the last twenty months and became a going geopolitical fact — and the clearest evidence of that is how differently it is going in two places that announced projects within months of each other.
Stargate UAE broke ground March 20, 2026: a $30 billion campus in Abu Dhabi's Masdar City, majority-owned by G42 (60%), with OpenAI holding 20% in exchange for AI models, training methodology, and research expertise, and Oracle, Nvidia, SoftBank, and Cisco as partners. The first 200-megawatt cluster is targeted live in 2026, scaling toward 5 gigawatts.16 Stargate UK, announced in September 2025 with Nvidia and Nscale — up to 31,000 GPUs at a site in North East England — was paused by OpenAI in April 2026, citing UK energy costs and regulatory uncertainty, with OpenAI stating it would “move forward when the right conditions…enable long-term infrastructure investment.”17 One project, offering direct majority equity control and effectively subsidized industrial power, broke ground within months. A comparable project in a G7 ally with a competing energy-price and regulatory regime stalled within seven months of announcement. “Sovereign AI” is not one policy environment. It is at least two, and they are moving in opposite directions.
The regulatory version of the same story is unfolding in Europe, and it turns on a distinction most coverage glosses over: jurisdiction of data, not jurisdiction of servers. The EU AI Act's high-risk-system provisions become fully applicable August 2, 2026; the European Commission formally proposed the Cloud and AI Development Act on June 3, 2026, establishing the EU's first four-tier cloud-sovereignty framework.25 The load-bearing legal fact underneath all of it is the U.S. CLOUD Act, which allows U.S. law enforcement to compel an American company to produce data it holds abroad — even if the servers physically sit in Frankfurt. Hosting in a hyperscaler's EU region does not, by itself, remove that exposure; only a genuinely EU-domiciled operator, or a Gaia-X Level 3-certified sovereign offering, does.25 AWS's response — a distinct “European Sovereign Cloud,” a legally separate EU entity with EU-only leadership and operational control — is a direct, dated concession that ordinary multi-region hosting was not sufficient.25 The global sovereign-cloud market is sized at $195.35 billion for 2026, a market-research estimate flagged as such, but directionally consistent with the regulatory activity documented above.25
Both threads — Stargate's uneven diplomacy and Europe's residency law — point at the same fact this layer keeps surfacing: the toll booth from Section 3 is not just a product decision inside a single company. It is increasingly a matter decided by which government you're standing in when you ask for the guarantee.
/ 06Follow the Equity, Follow the Token
Follow the equity: the cloud layer's headline equity story is that RPO announcements move stock prices independent of revenue or earnings — Oracle gained roughly a quarter-trillion dollars of market value on an RPO beat in a quarter it missed on both.8 That is not manufactured cynicism; it is the market's own revealed preference for a specific number. The bull case underneath it is genuinely real: AWS's cloud margin clears 37%, Azure grew 40%, and Microsoft's non-OpenAI commercial backlog alone grew 26% — a diversified, profitable, currently-collecting-cash business, not a mirage. The bear case is equally real and sits in the same filings: roughly half of the industry's $2.17 trillion backlog traces to two counterparties whose own revenue does not yet cover it, and the credit-rating ladder — hyperscaler investment-grade, to Oracle's negative-outlook investment-grade, to CoreWeave's and Nebius's 6%–15% debt — is the market's own priced acknowledgment of exactly how much of that backlog is promise versus payment.
Follow the token: the RPO circularity documented here is the cloud layer's version of Part 2's mark-to-market circularity, one rung up the stack. At the silicon layer, Nvidia's stakes in the labs got marked up as the subsidy; here, the backlog itself — a number with real accounting substance but limited cash certainty — is the vehicle. And the mechanism is no longer hidden in an accounting footnote. It is visible in named counterparties, in a Moody's downgrade rationale, in a CoreWeave risk factor, in a basis-point spread between Oracle's coupon and CoreWeave's. Somewhere near the middle of this circle, two names keep reappearing: OpenAI and Anthropic, whose combined compute commitments make up roughly half of everything this layer has booked.
So go up one more rung, to the labs that are simultaneously the demand this entire layer is built on and the counterparty risk its credit-rating agencies keep flagging by name. Part 4 follows the token into the frontier models — the companies whose $1.4 trillion of total commitments sit against $10–20 billion of current revenue, and whose next funding round is, in no small part, what makes this layer's backlog collectible at all.